Classification methods for neuronal axons and dendrites

By employing a multi-feature fusion and multi-level rule-based hierarchical decision-making method, the axons and dendrites of neurons are automatically classified, solving the problems of time-consuming, labor-intensive, and highly subjective methods in existing technologies, and achieving efficient and reliable neuronal morphological analysis.

CN123072009APending Publication Date: 2026-11-10HUST SUZHOU INST FOR BRAINMATICS
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611165859.7
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-03
Publication Date
2026-11-10

AI Technical Summary

Technical Problem

In existing technologies, classification methods for neuronal axons and dendrites are time-consuming, labor-intensive, highly subjective, and difficult to meet the needs of high-throughput data processing. Furthermore, automatic or semi-automatic methods are difficult to achieve precise classification.

Method used

A multi-feature fusion and multi-level rule-based hierarchical decision-making method is adopted. By obtaining the original SWC file, a tree topology map of neurons is generated, the main branch points are determined, and the confidence value propagation algorithm is used for refined classification, and finally the SWC file is updated.

Benefits of technology

It achieves efficient automatic classification of axons and dendrites with high consistency and repeatability, adapts to the morphological differences of neurons in different brain regions and types, significantly improves the overall efficiency of large-scale neuronal morphology analysis, and provides a quality assessment report.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123072009A_ABST
    Figure CN123072009A_ABST
Patent Text Reader

Abstract

This invention discloses a method for classifying neuronal axons and dendrites, comprising the following steps: S1, obtaining an original SWC file and generating a tree topology diagram of neurons based on the original SWC file; S2, determining a main branch point from the tree topology diagram of neurons; S3, determining the categories of the two main subtrees of the main branch point to obtain a preliminary classification result; S4, refining the preliminary classification result based on a confidence value propagation algorithm to obtain a final classification result; S5, updating the original SWC file according to the final classification result and outputting the updated SWC file. This invention achieves efficient automatic classification of axons and dendrites through multi-feature fusion and multi-level rule-based hierarchical decision-making. This method achieves accuracy comparable to experts while exhibiting high consistency and repeatability, and can adapt to the morphological differences of neurons in different brain regions and types, significantly improving the overall efficiency of large-scale neuronal morphological analysis.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of neuron image processing, and in particular to a method for classifying neuron axons and dendrites. Background Technology

[0002] Neurons are the basic structural and functional units of the nervous system, and their morphological characteristics are the core foundation for neuron type identification, neural circuit analysis, and brain function research. With the rapid development of microscopic optical imaging and digital reconstruction technologies, researchers have been able to acquire vast amounts of three-dimensional neuron morphological data and store and share them in the standardized SWC format. SWC files comprehensively describe the dendritic and axonal skeletal structure of neurons through a series of nodes with three-dimensional coordinates, radii, parent node connections, and type labels (such as cell body, axon, and dendrites). Currently, NeuroMorpho.org, the world's largest neuron morphology database, contains SWC reconstruction data for over 200,000 neurons, which are widely used in neuron classification, computational modeling, and brain mapping.

[0003] In a typical neuroscience research process, researchers first label specific neurons using techniques such as viral tracing and sparse labeling. Then, they acquire whole-brain image data using optical microscopy (e.g., fMOST, STPT). Finally, they use manual or semi-automatic reconstruction software (e.g., Neurolucida, Vaa3D) to trace nerve fibers point-by-point, generating the original SWC file. However, during reconstruction, operators typically only sample continuously along the fiber direction without functionally distinguishing and labeling axons and dendrites. Therefore, in the reconstructed SWC file, all neurite nodes are often temporarily labeled as the same type (e.g., type 0 or undefined), or simply as "dendrite." To meet subsequent morphological analysis needs (e.g., calculating the length ratio of axons to dendrites, analyzing differences in branching patterns, and determining the projection type of neurons), researchers must use specialized software, with experts possessing neuroanatomical knowledge, to manually subdivide and relabel the axons and dendrites of each neuron.

[0004] Currently, the most commonly used axon-dendritic segmentation method is manual segmentation. This method relies on professional neuronal morphology editing software (such as Neurolucida, Vaa3D, NeuronJ, neuTube, etc.) and the operator's neuroanatomical experience.

[0005] Taking the widely used open-source software Vaa3D as an example, the operator needs to load the target SWC file and the corresponding original microscopic image stack into the software to verify the continuity of neurites during the segmentation process. Using mouse dragging and zooming, the operator observes the overall morphology of the neuron from different angles, initially determining which processes are relatively short and densely branched (usually dendrites) and which are long and thin with a long extension (usually axons). Then, the operator searches for the first branching node near the cell body. For typical cell types such as cortical pyramidal neurons, axons typically originate from the base of the cell body or the opposite side of a major dendritic trunk, while dendrites are radially distributed around the cell body. The operator needs to repeatedly rotate the view to compare the morphological characteristics of each branch, including: (1) Direction characteristics: Axons tend to project to distant places (spanning several millimeters), while dendrites are mostly confined to a range of several hundred micrometers around the cell body.

[0006] (2) Thickness characteristics: Axons usually become thinner rapidly after the initial segment, while the diameter of dendrites decreases more gradually.

[0007] (3) Branching pattern: Dendritic branches have a large angle and are crown-shaped; axon branches have a small angle and are often parallel or bifurcated at a small angle.

[0008] (4) Spatial distribution: When a specific brain region has been marked, if a branch enters the white matter region, it is highly likely to be an axon. Based on the above characteristics, the most likely branch point is selected as the "axon-dendritic boundary point". Starting from the boundary point, a continuous segment of nodes and their corresponding branches are selected. The software's preset label options define which part is the dendritic part and which part is the axon part.

[0009] After making the modifications, save the file as a new file to complete the axonal dendrite division of a neuron reconstruction result.

[0010] The above-mentioned method of manually distinguishing axons and dendrites has significant drawbacks in practical applications, specifically in the following aspects: First, manual segmentation is extremely time-consuming and labor-intensive, making it difficult to meet the demands of high-throughput data processing. With advancements in automated imaging technology and reconstruction algorithms, the number of neuron reconstructions generated in a single research project has increased from dozens to tens of thousands. For each neuron, experts need to start from the cell body and repeatedly compare and trace along the fiber paths of dendrites and axons to determine its orientation and morphological characteristics before identifying and marking the dividing points. Completing a moderately complex neuron segmentation typically takes several minutes to tens of minutes. Faced with thousands of neurons, even with multiple experts dedicating their full-time efforts, it would still take months, which has become a bottleneck restricting large-scale neuronal morphological research.

[0011] Secondly, the subjective nature of manual classification leads to inconsistent data standards and poor reproducibility. Different experts have different criteria for judging axons and dendrites: some rely more on fiber thickness and branching angles, while others focus more on projection distance and endpoint regions. Even the same expert may have discrepancies in classifying the same neuron at different time points. This subjectivity makes it difficult to integrate and compare data generated by different laboratories, or even within the same laboratory, across experiments and batches, severely impacting data standardization and reuse value.

[0012] Furthermore, manual segmentation demands a high level of expertise from the operator, hindering the widespread adoption of the technology. Accurately distinguishing axons and dendrites requires solid neuroanatomical knowledge and extensive practical experience, posing a significant hurdle for novice graduate students or researchers outside the field. This reliance on an apprenticeship-style knowledge transfer method is detrimental to the large-scale promotion and widespread application of this technology.

[0013] Finally, existing automated or semi-automated methods struggle to directly address this problem. While some deep learning-based methods classify neurons (e.g., determining whether they are "pyramidal cells" or "basket cells"), these methods focus on overall type identification rather than fine-grained classification of the internal structure of individual neurons. Furthermore, there are tools for validating and converting SWC files, but they lack the biological ability to distinguish between axons and dendrites. Summary of the Invention

[0014] The purpose of this invention is to solve the above-mentioned problems existing in the prior art and to provide a method for classifying neuronal axons and dendrites.

[0015] The objective of this invention is achieved through the following technical solution: The classification method for neuronal axons and dendrites includes the following steps: S1, Obtain the original SWC file, and generate a tree topology diagram of neurons based on the original SWC file; S2, determine a main branch point from the tree topology of the neuron; S3, determine the categories of the two main subtrees of the main branch point to obtain preliminary classification results; S4, refine the preliminary classification results based on the confidence value propagation algorithm to obtain the final classification results; S5 updates the original SWC file based on the final classification results and outputs the updated SWC file.

[0016] Preferably, S2 includes: S21, determine candidate branch points based on the tree topology diagram; S22, determine the subtree asymmetry score, spatial separation score, topological distance score and morphological difference score for each candidate branch point; S23. Calculate the comprehensive score of each candidate branch point based on the subtree asymmetry score, spatial separation score, topological distance score and morphological difference score of each candidate branch point. S24, the candidate branch with the highest comprehensive score is taken as the main branch.

[0017] Preferably, in step S23, the comprehensive score of each candidate branch point is obtained by weighted summation of its subtree asymmetry score, spatial separation score, topological distance score, and morphological difference score.

[0018] Preferably, when determining the weights of the subtree asymmetry score, spatial separation score, topological distance score, and morphological difference score for each candidate branch point, the global morphological features of the tree topology are first determined, and then it is determined whether the weights need to be adjusted based on the global morphological features. If so, the weights are adjusted according to a preset rule; otherwise, the default weights are used.

[0019] Preferably, the default weights corresponding to the subtree asymmetry score, spatial separation score, topological distance score, and morphological difference score are 0.4, 0.3, 0.2, and 0.1, respectively.

[0020] Preferably, the global morphological features of the tree topology graph include global branch ratio, global spatial extensibility, and global maximum topological depth.

[0021] Preferably, S3 includes: S31, divide all direct subtrees of the main branch point into two mutually exclusive main subtrees; S32, determine the distal scale feature value, density feature value and orientation feature value of the two main subtrees respectively; S33, determine whether the absolute value of the difference between the distal scale feature values ​​of the two main subtrees exceeds a first set value; if yes, determine the main subtree with the larger distal scale feature value as an axon and the main subtree with the smaller distal scale feature value as a dendrite; if no, execute S34. S34, determine whether the absolute value of the difference between the density feature values ​​of the two main subtrees exceeds the second set value. If yes, determine the main subtree with the larger density feature value as a dendrite and the main subtree with the smaller density feature value as an axon. If no, proceed to S35. S35, determine whether the absolute value of the difference between the directional feature values ​​of the two main subtrees exceeds the third preset value. If yes, determine the main subtree with the higher directional feature value as an axon and the main subtree with the lower directional feature value as a dendrite. If no, proceed to S36. S36, the main subtree with more nodes in the two main subtrees is identified as an axon, and the main subtree with fewer nodes is identified as a dendrite.

[0022] Preferably, S4 includes: S41, perform anomaly detection on the preliminary classification results, and correct anomalies when they are detected; S42, assign initial confidence values ​​to the main branch point and all its downstream nodes; S43, Update the confidence values ​​of the main branch point and all its downstream nodes; S44, perform reclassification and update the confidence value for boundary nodes located on the axon and dendrite boundaries and whose confidence values ​​are lower than the preset threshold; S45, iterate through S43-S44 until the confidence values ​​of all nodes downstream of the main branch point converge, and obtain the final classification result.

[0023] Preferably, in step S43, the confidence value of each node is updated according to the following formula: The updated node confidence value = 0.7 × the current node confidence value + 0.3 × the support ratio of neighbors with the same label.

[0024] Preferably, a quality assessment report is also output in step S5.

[0025] The advantages of the technical solution of this invention are mainly reflected in: This invention achieves efficient automatic classification of axons and dendrites through multi-feature fusion and multi-level rule-based hierarchical decision-making. While achieving accuracy comparable to expert methods, this approach exhibits high consistency and repeatability, and can adapt to the morphological differences of neurons in different brain regions and types, significantly improving the overall efficiency of large-scale neuronal morphology analysis. Furthermore, this invention offers good interpretability and can automatically generate effective quality assessment reports, providing users with a quantitative assessment of the reliability of classification results and facilitating the identification of low-confidence regions that may require manual review. Attached Figure Description

[0026] Figure 1 This is a schematic diagram of the overall process of the present invention; Figure 2 This is a flowchart of step S2 of the present invention; Figure 3 This is a flowchart of step S3 of the present invention; Figure 4 This is a flowchart of step S4 of the present invention. Detailed Implementation

[0027] The objectives, advantages, and features of this invention will be illustrated and explained through the following non-limiting description of preferred embodiments. These embodiments are merely typical examples of applying the technical solutions of this invention, and all technical solutions formed by equivalent substitutions or equivalent transformations fall within the scope of protection claimed by this invention.

[0028] In the description of the solution, it should be noted that the terms "center," "upper," "lower," "left," "right," "front," "rear," "vertical," "horizontal," "inner," and "outer," etc., indicating the orientation or positional relationship, are based on the orientation or positional relationship shown in the accompanying drawings and are only for the convenience and simplification of description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the present invention. Furthermore, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0029] Example 1 The classification method for neuronal axons and dendrites disclosed in this invention will be described below with reference to the accompanying drawings. Figure 1 As shown, it includes the following steps: S1, Obtain the original SWC file, and generate a tree topology diagram of neurons based on the original SWC file; S2, determine a main branch point from the tree topology of the neuron; S3, determine the categories of the two main subtrees of the main branch point to obtain preliminary classification results; S4, refine the preliminary classification results based on the confidence value propagation algorithm to obtain the final classification results; S5 updates the original SWC file based on the final classification results and outputs the updated SWC file.

[0030] Specifically, in S1, the original SWC file is a reconstructed SWC file. The SWC format is a standard file format widely used in neuroscience to store the three-dimensional morphological data of a single neuron in a structured manner. A standard SWC record line represents a neuron morphological sampling point (node), which includes the following seven types of data: node number, structure type, X coordinate, Y coordinate, Z coordinate, radius, and parent node number.

[0031] In S1, the core of constructing the tree topology graph of neurons lies in parsing the seven types of data of each node, establishing the mapping relationship between parent nodes and child nodes in memory, and forming a tree topology graph. The specific process of constructing the tree topology graph of neurons is a well-known technology in the field and is not an innovation of this invention, so it will not be described in detail here.

[0032] After constructing the tree topology graph, before executing S2, an integrity check can be performed on the constructed tree topology graph, including at least the following checks: (1) Whether there is a loop, that is, whether there is a situation where you can return to the same node by starting from the same node and following the parent-child relationship. If so, it means that there is a loop.

[0033] (2) Whether the number of connected components is 1, that is, whether it is possible to access all other nodes starting from the unique root node. If yes, the number of connected components is 1; otherwise, the number of connected components is greater than 1.

[0034] (3) Whether there are isolated nodes, that is, whether there are nodes that have no parent-child relationship. If so, it means that there are isolated nodes.

[0035] (4) Whether the parent node reference is valid, that is, whether the parent node of each node points to an actual existing node. If so, the parent node reference is valid. (5) Whether there are duplicate coordinates, that is, whether there is a three-dimensional spatial coordinate corresponding to multiple nodes. If so, it means that there are duplicate coordinates. (6) Whether there is a self-loop, that is, the parent node of a certain node is itself. If so, it means that there is a self-loop.

[0036] (7) Check whether the coordinates and radius values ​​are valid, that is, check whether the coordinates and radii of all nodes are reasonable finite values. For example, they cannot be NaN (not a number), Inf (infinity) or negative numbers (the radius cannot be negative).

[0037] The specific methods for integrity checks are well-known techniques in the field and are not innovative in this invention, so they will not be described in detail here.

[0038] When the integrity check finds fatal topology errors such as multi-connected components, loops, or invalid parent nodes, the above-mentioned anomalies are output in the form of a check report. At the same time, the subsequent process is stopped, and manual review is prompted as an option.

[0039] When the integrity check meets the requirements, that is, when the tree topology graph satisfies the requirements of simple connectivity, no loops, no invalid parent nodes, and valid coordinate and radius values, the root node of the tree topology graph is automatically relocated to a cell node. The specific steps include the following: Cell node identification: Cell nodes are identified by scanning the Type attribute of all nodes in the original SWC file. Among the nodes with Type=1 (the label corresponding to the cell), the one with the largest radius is selected as the cell node; if there is no label, the node with the largest radius is selected as the cell node.

[0040] Reroot: Perform a breadth-first traversal (BFS) with the identified cell node as the root node, assign a consistent direction outward from the cell node to each edge in the traversal order, and update the parent pointers of each node.

[0041] S2 includes the following steps: S21, candidate branch points are determined according to the tree topology graph, and each candidate branch point is a node in the tree topology graph with an out-degree greater than or equal to 2; S22, determine the subtree asymmetry score, spatial separation score, topological distance score, and morphological difference score for each candidate branch point.

[0042] (1) The subtree asymmetry score S asym This method measures the difference in the number of nodes between the largest and second-largest direct subtrees under a candidate branch point. A greater difference indicates that the candidate branch point is more likely to be an axon-dendritic boundary. The largest direct subtree is the subtree downstream of the candidate branch point with the most leaf nodes, and the second-largest direct subtree is the direct subtree downstream of the candidate branch point with the second-largest number of leaf nodes. The number of leaf nodes is closer to the biological distinguishing goal, as axons and dendrites typically exhibit two different types of terminal branching systems of different sizes and complexities on either side of the main boundary. Leaf nodes are the true ends (out-degree of 0) in the tree topology, and their number directly corresponds to how many independent terminal paths are ultimately formed by the direct subtree. When locating candidate branch points that "can cut a neuron into two major functional compartments," the focus is on whether two sufficiently large terminal systems are formed on both sides, rather than how many sampling points are taken along the middle path. Therefore, defining the largest / second-largest direct subtree using the number of leaf nodes is consistent with the task goal of "finding the axon-dendritic boundary."

[0043] The total number of nodes in a direct subtree is easily affected by the sampling density of the SWC reconstruction, resulting in poor robustness. The original SWC often samples densely or sparsely along the fiber direction: on the same real cable segment, overly dense reconstruction produces a large number of non-leaf intermediate nodes, significantly inflating the "total number of nodes"; overly sparse reconstruction results in fewer nodes. Therefore, the total number of nodes is influenced by the operator's / algorithm's sampling habits and may not reflect the true size and shape. In contrast, the number of leaf nodes is mainly determined by the tree's branch topology, making it less sensitive to oversampling / undersampling along the edge direction, and allowing for comparisons across datasets and reconstruction batches. Furthermore, using the number of leaf nodes facilitates filtering out noisy short branches, improving the quality of candidate main branch points.

[0044] The subtree asymmetry score S asym The specific determination process is as follows: Get a list of leaf nodes in each direct subtree of a candidate branch point; Filter out direct subtrees with leaf node count Li≤2; Define the number of leaf nodes in the largest and second-largest direct subtrees as L. MAX L 2nd ; The intermediate value raw is calculated using the following formula: raw=L MAX ×L 2nd ; The intermediate value raw is normalized according to the following formula to obtain the normalized value norm: ; in, This is the sum of the number of leaf nodes in all direct subtrees retained after filtering.

[0045] The asymmetric proportion r is calculated using the following formula. asym : r asym =L MAX / (L MAX +L 2nd ); The subtree asymmetry score S of the candidate branch point is determined according to the following formula. asym : S asym =norm×r asym ∈[0,1].

[0046] (2) The spatial separation score S sep It is determined by calculating the cosine of the angle between the centroid directions of the largest and second largest direct subtrees of the candidate branch point. The greater the spatial separation, the more obvious the difference in the extension directions of the largest and second largest direct subtrees in three-dimensional space.

[0047] The specific calculation process is as follows: The largest direct subtree is represented by C1, the second largest direct subtree by C2, and C2 has more than 2 leaf nodes. If the number of leaf nodes in C2 is less than or equal to 2, then the spatial separation degree S is directly determined. sep =0.

[0048] Define the node sets V1 and V2 of C1 and C2 as follows: V1 = Desc(C1) ∪ C1; V2=Desc(C2)∪C2; Where Desc() represents all descendant nodes downstream of a direct subtree; The three-dimensional coordinates of any node p in C1 and C2 are C p =(xp, yp, zp); Calculate the arithmetic mean of the coordinates of all nodes in C1 and C2 respectively, and use it as the coordinates of the centroids of C1 and C2. The specific calculation formula is as follows; ; ; Where |V1| and |V2| are the cardinality of the node sets V1 and V2 of C1 and C2, respectively, that is, the number of nodes contained in each node set V1 and V2. p The coordinates of the p-th node in the largest or second-largest direct subtree; Calculate the value from candidate branch point bp (coordinate C) bp The direction vectors v1 and v2 pointing to the centroids of C1 and C2 are: v1=C1 质 -C bp ; v2=C2 质 -C bp ; Calculate the cosine of the included angle using two direction vectors v1 and v2 and the following formula: cosθ=(v1·v2) / ( ‖v1‖×‖v2‖); When the magnitudes of the direction vectors, |v1| and / or |v2|, are less than a set value (which can be customized as needed, but is not limited here), S is directly determined. sep =0; Conversely, the spatial separation score S can be determined according to the following formula. sep : S sep =(1-cosθ) / 2; When cosθ=1, S sep =0; When cosθ=0, S sep =0.5; When cosθ = -1, S sep =1.

[0049] (3) The topological distance score S dist The scoring is based on a Gaussian scoring function and the ratio of the topological distance from each candidate branch point to the cell node (root node) to the total path length. The peak value is set at 30%, reflecting the prior knowledge that the main branch point usually appears at a certain distance from the cell node.

[0050] The topological distance score S dist Determined according to the following process: Calculate the number of edges d in the shortest path from each candidate branch point bp to the cell body node; The number of edges d in the shortest path from each candidate branch point bp to the cell body node is normalized according to the following formula: d'=(d-dmin) / (dmax-dmin); Where d' is the normalized value, dmax is the maximum number of edges in the path from all candidate nodes to the cell body node, and dmin is the minimum number of edges in the shortest path from all candidate nodes to the cell body node.

[0051] The topological distance score S is determined according to the following Gaussian scoring formula. dist : S dist =exp(−(d'−0.3) 2 / (2×0.22)); Here, exp() represents the exponential operation with the natural constant e as the base.

[0052] When dmax = dmin, then S dist =0.5.

[0053] (4) The morphological difference score S morph It comprehensively compares the differences between the largest and second largest direct subtrees under candidate branch points in three aspects: subtree branch density, average subtree side length, and subtree topological depth, capturing the structural differences between axons and dendrites in micromorphology.

[0054] The morphological difference score S morph Determined according to the following process: Calculate the subtree branch density ρ of the largest and second largest direct subtrees under the candidate branch points, respectively. 子 The average side length of the subtree is ē 子 and the maximum topological depth D of the subtree 子 ; Wherein, the subtree branch density ρ 子 =Number of branches in the direct subtree / Total number of nodes in the direct subtree; Average side length of the subtree 子 It is the average length of all edges in the direct subtree; the maximum topological depth D of the subtree. 子 It is the maximum number of edges (hops) traversed from the root node of the direct subtree to each leaf node in the direct subtree. For example, the maximum topological depth of the subtree of the maximum direct subtree is the maximum number of edges (hops) traversed from the root node (direct child node of the candidate branch point) of the maximum direct subtree to each leaf node in the maximum direct subtree.

[0055] Calculate the branch density ρ of the largest and second-largest direct subtrees using the following formulas. 子 Average side length of subtrees 子 and the maximum topological depth D of the subtree 子 The relative difference Δ f ; Δ f =|f1−f2| / (f1+f2); Where f1 represents the subtree branch density, average subtree side length, or maximum topological depth of the largest direct subtree; and f2 represents the subtree branch density, average subtree side length, or maximum topological depth of the second largest direct subtree. For example, calculate the relative difference Δ between the subtree branch densities of the largest and second largest direct subtrees. ρ子 When f1 is the subtree branch density of the largest direct subtree, f2 is the subtree branch density of the second largest direct subtree.

[0056] Determine the relative difference Δ between the branch densities of the largest and second-largest direct subtrees. ρ子 The relative difference Δ in the average side length of subtrees ē子 and the relative difference Δ between the maximum topological depth of subtrees D子 The average value is used as the morphological difference score S. morph S morph =(Δ ρ子 +Δ ē子 +Δ D子 ) / 3.

[0057] S23, calculate the comprehensive score for each candidate branch point based on its subtree asymmetry score, spatial separation score, topological distance score, and morphological difference score; specifically, the comprehensive score S for each candidate branch point is... bp It is obtained by weighted summation of its subtree asymmetry score, spatial separation score, topological distance score, and morphological difference score, i.e., determined according to the following formula: S bp =w a ×S asym +w s ×S sep +w d ×S dist +w m ×S morph ; Among them, S bp S is the comprehensive score of the candidate branch points. asym Score the asymmetry of the subtree; S sep Score the spatial separation. dist S is the topological distance score; morph Scoring for morphological differences; w a w s w d w m For weights.

[0058] When determining the weights of subtree asymmetry score, spatial separation score, topological distance score, and morphological difference score for each candidate branch point, the global morphological features of the tree topology are first determined, and it is determined whether the weights need to be adjusted based on the global morphological features of the tree topology. If so, the weights are adjusted according to preset rules; otherwise, the default weights are used.

[0059] The global morphological features of the tree topology include: (1) Global branching ratio Ratio It can be determined according to the following formula: branching Ratio =N bp / N 总 ; Where, N bp N represents the number of branch nodes in the tree topology graph. 总 This represents the total number of nodes in the tree topology graph. (2) Global spatial stretch Extent It can be determined according to the following formula: spatial Extent =‖bbox max -bbox min ‖ / ē edge ; ē edge =L total / E; Where bbox is the bounding box of the three-dimensional coordinates of all nodes in the tree topology graph, ||bbox max -bbox min ‖ represents the diagonal length of the bounding box, ē edge L is the average side length. total The total cable length of the tree topology is the sum of the Euclidean lengths of all edges in the tree topology, and E is the number of edges in the tree topology.

[0060] (3) Global maximum topology depth D topo,max It is the maximum number of edges required to reach each leaf node of the tree topology from the cell node.

[0061] After determining the aforementioned global morphological features, it is determined whether weight adjustments are needed, and if so, how to adjust them according to preset rules. Specifically, when adjusting the weights for the subtree asymmetry score, spatial separation score, topological distance score, and morphological difference score, adjustments are made in conjunction with the morphological factor α (which can be customized, for example, 0.2). If spatial Extent>150, adjustment is needed. Adjustment rules: w s Increase, w a Decrease, w s The increase is equal to a × 0.1; w a The decrease is equal to a × 0.05; If branching Ratio >0.15, adjustment is needed. Adjustment rules: w a w m Increase, w a The increase is equal to a × 0.1, w m The increase is equal to a × 0.05; If D topo,max >50, adjustment is needed. Adjustment rules: w m Increase, w d Decrease, w m The increase is equal to a × 0.1, w d The decrease is equal to a × 0.05; After adjusting the weights according to at least one of the above adjustment rules, normalization is performed on the adjusted weights so that the sum of the four is 1, resulting in the final updated weights.

[0062] Conversely, if none of the above conditions are met, i.e., spatial Extent ≤150 and branching Ratio ≤0.15 and D topo,max If the value is ≤50, the default weights are used. The default weights for the subtree asymmetry score, spatial separation score, topological distance score, and morphological difference score are 0.4, 0.3, 0.2, and 0.1, respectively.

[0063] S24, the candidate branch point with the highest comprehensive score is selected as the main branch point. The main branch point theoretically divides the neuron into two main parts: the axon and the dendrite.

[0064] After identifying the main branch point, the two main subtrees downstream of it are selected as candidate axons and dendrites and classified through a multi-level waterfall decision-making mechanism. The multi-level waterfall decision-making mechanism is based on the design principle of "prioritizing strong signals, supplementing weak signals step by step, and using a fallback strategy as a fallback".

[0065] Correspondingly, S3 includes the following steps: S31, divide all direct subtrees of the main branch point into two mutually exclusive main subtrees.

[0066] At the determined main branch point mp, all direct subtrees of the main branch point mp are divided into two mutually exclusive main subtrees T1 and T2 (there are only two main subtrees). The specific process is as follows: (1) Arrange the direct subtrees of the main branch point mp in descending order of the number of nodes; (2) The main subtree T1 = the direct subtree with the most nodes; (3) The main subtree T2 = the remaining direct subtrees excluding the direct subtree with the most nodes.

[0067] For example, after the main branch point mp, there are four direct subtrees C1, C2, C3, and C4, with the total number of nodes in the four direct subtrees being 120, 70, 55, and 30 respectively. At this time, T1 = C1 and T2 = C2∪C3∪C4, that is, T1 has 120 nodes and T2 has 155 nodes.

[0068] S32 extracts the multidimensional feature vectors of the two main subtrees and determines the far-end scale feature value, density feature value, and orientation feature value of the two main subtrees respectively, providing a comprehensive quantitative basis for hierarchical decision-making.

[0069] Specifically, for the main subtrees T1 and T2 obtained from the division, their respective branch features are extracted. The specific branch features include, but are not limited to: the maximum topological depth of the main subtree, the average topological depth of the main subtree, the main axis length of the bounding box of the main subtree, the branch point density of the main subtree, the leaf node ratio of the main subtree, and the node density of the main subtree.

[0070] The subtree topological depth is determined by taking the starting point s of the main subtree Ti (i=1 or 2) as the root and performing a breadth-first traversal along the directed edges between the parent and child trees within the main subtree. The depth is increased by 1 for each edge traversed.

[0071] Let the set of leaf nodes be L = {n ∈ Ti | outdeg(n) = 0}, where outdeg(n) represents the out-degree of node n; then the maximum topological depth d of the primary subtree. topo,max =max l∈L depth(l) represents the maximum number of edges required to reach each leaf node of the main subtree from the starting point s; the average topological depth d of the main subtree topo,mean =mean l∈L depth(l) represents the average number of edges required to reach each leaf node of the main subtree from the starting point s.

[0072] The main subtree bounding box main axis length b box,major It takes the maximum value among the three-axis side lengths of the bounding box formed by the minimum and maximum values ​​of the three-dimensional coordinates of all nodes within a main subtree. For example, if the three-axis side lengths of the bounding box are (Δx, Δy, Δz), then b box,major =max(Δx, Δy, Δz), bounding box volume V bbox =Δx·Δy·Δz, if V bbox <10-10 Then take V bbox =10 -10 .

[0073] The main subtree branch point density Branching Ratio Determined according to the following formula: Branching Ratio =N 主 / |Ti|; Where, N 主 Ti represents the number of nodes with an out-degree ≥ 2 in the main subtree, and |Ti| represents the total number of nodes in the main subtree.

[0074] The proportion of leaf nodes of the main sub-trees r leaf Determined according to the following formula: r leaf =N leaf / |Ti|; Where, N leaf This represents the number of nodes (leaf nodes) with an out-degree of 0 within the main subtree.

[0075] The density ρ of the main subtree nodes 主 =|Ti| / V bbox ; Define the main subtree principal axis direction vector m=p lfar -p soma , where p lfar The leaf node l that is furthest from the cell body node in the main subtree. far The three-dimensional coordinates, p soma represents the three-dimensional coordinates of the cell body node. It indicates the main direction in which the main subtree extends outward from the cell body node.

[0076] If the coordinates c of the centroid of the main subtree are the arithmetic mean of the coordinates of all nodes within the main subtree, then the centroid vector v = cp. soma .

[0077] The cosine of the angle between the centroid vector and the direction vector of the principal axis of the main subtree, cosβ, is calculated according to the following formula, with the value pruned to [-1, 1]: cosβ=(v⋅m) / (‖v‖×‖m‖); Where, ||v|| and ||m|| are the magnitudes of vectors v and m.

[0078] Based on the branch characteristics of the two main subtrees described above, determine the distal scale feature values ​​S of the two main subtrees respectively. far Density eigenvalue S dense , directional eigenvalue S dir .

[0079] (1) The distal scale feature value S far Taking into account factors such as the topological depth, geometric distance, and bounding box range of the main subtrees, the specific determination is based on the following formula: ; Where, d topo,max The maximum topological depth of the main subtree; D topo,max The maximum topological depth of the tree topology graph; d geo,max It is the maximum geometric path length within the main subtree, that is, the maximum value of the geometric path length (the sum of the Euclidean distances of each edge of the path) of the path from the starting point s to each leaf node of the main subtree. D geo,max The maximum geometric path length of the tree topology is the maximum value among the geometric path lengths of the paths from the cell node to each leaf node in the tree topology. d geo,mean The average geometric path length within the main subtree is the average of the geometric path lengths from the starting point s of the main subtree to each leaf node of the main subtree. D geo,mean The average geometric path length of the tree topology is the average geometric path length of the path from the cell node to each leaf node in the tree topology. b box,major This is the maximum value among the three side lengths of the bounding box; B max The maximum value among the three-axis side lengths of the bounding box for aligning the coordinate axes of all nodes in the tree topology graph is used. The specific method for determining the maximum value among the three-axis side lengths is a well-known technique and is not an innovation of this invention, so it will not be elaborated here.

[0080] (2) The density characteristic value S dense Determined according to the following formula: S dense =0.5×Branching Ratio +0.3×r leaf +0.2×ρ norm ; ρ norm =(ρ 主 -ρ min ) / (ρ max -ρ min ); Among them, Branching Ratio r is the density of branch points in the main subtree. leaf The proportion of leaf nodes in the primary subtree; ρ 主ρ represents the density of nodes in the primary subtree. min ρ max These represent the minimum and maximum values ​​of the density of the main subtree nodes of T1 and T2.

[0081] (3) The directional characteristic value S dir Determined according to the following formula: S dir =0.6×(‖v‖ / D max )+0.4×(1-cosβ); Where, ||v|| is the magnitude of the centroid of the main subtree relative to the centroid vector v of the cell body node; D max D is the maximum Euclidean distance from all nodes to the cell body nodes in the tree topology graph. In degenerate cases such as neurons containing only a single node, D... max Set 1.0 as the default value; cosβ is the cosine of the angle between the centroid vector and the direction vector of the main subtree's principal axis.

[0082] S33, First-level decision – Distant-scale analysis: Based on the determined distal-scale feature values ​​of the two main subtrees, determine whether the absolute value of the difference between the distal-scale feature values ​​of the two main subtrees exceeds a first preset value. If so, determine the main subtree with the larger distal-scale feature value as an axon and the main subtree with the smaller distal-scale feature value as a dendrite. Assign the corresponding axon label to all nodes of the main subtree determined to be an axon, and the corresponding dendrite label to all nodes of the main subtree determined to be a dendrite. If not, proceed to S34. The first preset value can be set as needed, preferably 0.25.

[0083] S34, Secondary Decision – Density Analysis: Based on the determined density characteristic values ​​of the two main subtrees, determine whether the absolute value of the difference between the density characteristic values ​​of the two main subtrees exceeds a second preset value. If yes, determine that the main subtree with the larger density characteristic value is a dendrite, and the main subtree with the smaller density characteristic value is an axon; if not, proceed to S35. The second preset value can be set as needed, preferably 0.05.

[0084] S35, Level 3 decision-making – direction analysis: Based on the directional feature values ​​of the two determined main subtrees, determine whether the absolute value of the difference between the two directional feature values ​​exceeds a third set value. If so, the main subtree with the higher directional feature value is determined as an axon, and the other main subtree is determined as a dendrite. If not, proceed to S36. The third set value can be set as needed, preferably 0.1.

[0085] S36, Backtracking Decision: The main subtree with more nodes is designated as the axon, and the main subtree with fewer nodes is designated as the dendrite. This strategy is based on the biological prior knowledge that axons typically have longer extension distances and more nodes.

[0086] After the above multi-level waterfall decision-making mechanism, the preliminary classification results are output (the labels of the axons or dendrites corresponding to each node). At the same time, the decision level that made the judgment is recorded (for example, the preliminary classification result is derived from the first-level decision) and the score difference between the two main subtrees on which the decision level is based (for example, the score difference of the first-level decision is the absolute value of the difference between the two distal scale feature values), which provides a reference for the iterative refinement in the subsequent step S4.

[0087] Step S4 refines the preliminary classification results generated in step S3 to improve the accuracy of classification boundaries and the smoothness of the overall results. Specifically, it includes the following steps: S41, morphological anomaly detection is performed on the preliminary classification results to identify potential isolated nodes (labels inconsistent with all topological neighbors), small fragments (insufficient number of connected component nodes within the same compartment), abnormal compartment topology (e.g., axonal disconnection), and excessive axon-dendritic spatial overlap. Detected anomalies are automatically repaired and / or manually reviewed to correct them. For example, isolated points can be automatically corrected to match the label of their parent node. For anomalies that cannot be automatically repaired, they are recorded and sent to a manual review queue, where human experts (e.g., neuroscientists) correct them by dragging and dropping, manually cutting or merging paths. The compartments are divided into axonal compartments and dendritic compartments. The set of all nodes currently labeled as axons is defined as an axonal compartment, and the set of all nodes currently labeled as dendrites is defined as a dendritic compartment. Cell body nodes are not compartments.

[0088] S42 assigns initial confidence values ​​to the main branch point and its downstream nodes; taking into account the quality score of the main branch point and the distance decay factor from each node to the main branch point, the node closer to the main branch point has a higher initial confidence value.

[0089] The initial confidence value (base) of the main branch point is determined according to the following formula: base = 0.5 + 0.3 × Q mbp Among them, Q mbp The quality score for the identified main branch points; Q mbp =0.4×Q scroe +0.3×Q spatial +0.2×Q topo +0.1×Q local .

[0090] Among them, Q spatial For space quality rating, Q topo Scoring for differences in topological rationality, Q local For local geometric quality scoring, Qscroe Basic discrimination score; (1) The Q spatial Determined according to the following process: Calculate the centroid distance d between the largest and second-largest direct subtrees A and B of the main branch point using the following formula. cent ; d cent =‖C A -C B ||; Among them, C A Let C be the centroid coordinates of the largest direct subtree A. B Let C be the centroid coordinates of the largest direct subtree B; the centroid coordinates are obtained by calculating the geometric mean of the three-dimensional coordinates of all nodes within a subtree. A -C B || represents finding vector C A -C B The length of the module.

[0091] The centroid distance is normalized according to the following formula: d norm =d cent / max((extent A +extent B ) / 2,ε); Among them, extent A extent B The diagonal length of the bounding box is determined based on the coordinates of all nodes in the largest and second-largest direct subtrees. The specific calculation method is a well-known technique and is not an innovation of this invention, so it will not be elaborated here.

[0092] Calculate the distance score Q using the following formula. dist : Q dist =tanh(d norm )∈[0,1]; Where tanh() represents the hyperbolic tangent function; Calculate the principal direction angle γ using the following formula: γ = arccos(|VA·VB|); Where VA is the direction vector of the largest direct subtree A, and VB is the direction vector of the second largest direct subtree B; the direction vector of a direct subtree can be obtained by performing PCA (principal component analysis) on the coordinates of all nodes of the subtree. The specific determination process is a well-known technique and is not an innovation of this invention, so it will not be elaborated here. |VA·VB| represents the absolute value of the dot product (inner product) result of vectors VA and VB.

[0093] Calculate the angle score Q using the following formula. angle ; Q angle =γ / 180∈[0,1]; Then, the space quality score Q spatial The following formula is used for calculation: Q spatial =0.6×Q dist +0.4×Q angle .

[0094] (2) The topological rationality difference score Q topo Determined according to the following process: Calculate the subtree branch density, subtree path depth, and number of leaf nodes in the largest direct subtree A and the second largest direct subtree B, respectively; subtree path depth L path L represents the maximum value among the total number of downstream descendants of all nodes within a subtree. Specifically, this can be determined by first identifying the total number of downstream descendants of each node within the subtree, and then selecting the maximum value from these totals as the subtree path depth L. path .

[0095] The differences in subtree branch density Δρ, subtree path depth ΔL, and subtree leaf node count ΔN between the largest direct subtree A and the second largest direct subtree B are calculated using the following formulas. leaf : Δρ=|ρ A -ρ B |; ΔL=|L pathA -L pathB | / (|A|+|B|); ΔN leaf =|N leaf,A -N leaf,B | / (N leaf,A +N leaf,B ); Where, ρ A ρ B The subtree branch densities of A and B are respectively, |ρ A -ρ B | represents the absolute value of the difference in subtree branch densities between A and B; L pathA L pathB The subtree path depths of A and B are respectively, |L pathA -L pathB | represents the absolute value of the difference in path depth between the subtrees of A and B; |A| and |B| represent the total number of nodes in A and B, respectively; N leaf,A N leaf,B Let N be the number of leaf nodes in the subtrees of A and B respectively. leaf,A -Nleaf,B | represents the absolute value of the difference between the number of leaf nodes in the subtrees of A and B.

[0096] The topological rationality difference score Q is determined according to the following formula. topo Q topo =(Δρ+ΔL+Δ Nleaf ) / 3.

[0097] (3) The local set quality score Q local The body is determined according to the following steps: The following local features are extracted within a neighborhood of radius R (e.g., 20 μm) centered on the main branch point: 1) Local branch density ρ local =Number of branch points in the neighborhood / Total number of nodes in the neighborhood; 2) Local radius change rate δᵣ=σ(r) / μ(r), where r: represents the radius value of each node in the neighborhood, σ(r) is the standard deviation of the radii of all nodes in the neighborhood, and μ(r) is the arithmetic mean of the radii of all nodes in the neighborhood.

[0098] 3) Neighborhood complexity k = number of neighborhood nodes / neighborhood volume; The local geometric quality score Q local Determined according to the following formula: Q local =0.5×S branch +0.3×S radius +0.2×S complex ∈[0,1]; S branch =min(5×ρ local ,1); S radius =min(2×δᵣ,1); S complex =exp(-|k-k0| / k0), k0=1×10 -4 .

[0099] (4) The basic discrimination Q scroe Determined according to the following formula: Q scroe =0.7×S bp 1 / (S bp 1 +1)+0.3×[(S bp 1 -S bp 2 ) / S bp 1 ]; Among them, Sbp 1 The overall score of the main branch (the candidate branch with the highest overall score), S bp 2 The overall score is the second highest among the candidate branches.

[0100] The initial confidence value of the cell body node is fixed at 1.0; Initial confidence value of nodes downstream of the main branch point (node) Determined according to the following formula: Conf (node) =max(0.3, base×exp(-d (node,mbp) / 50)); Where exp() represents exponential operation with the natural constant e as the base; d (node,mbp) This represents the length of the topological shortest path from a node downstream of the main branch point mp to the main branch point mp, which is the number of edges in the topological shortest path.

[0101] S43, update the confidence value of the main branch point and each downstream node; specifically, update the confidence value of each node according to the following formula: The updated node confidence value = 0.7 × the node's current confidence value + 0.3 × the proportion of neighbors with the same label; The same-label neighbor support ratio R same Determined according to the following formula: ; ; N(u) represents the neighboring nodes (parent / child, excluding cell bodies) of node u. Nsame(u) represents the neighboring nodes of node u that have the same label as node u. d(u,v): The Euclidean distance between node u and its neighbor node v; Conf(v): The current confidence value of neighbor node v.

[0102] S44, perform reclassification on boundary nodes located on the axon and dendrite boundaries and whose confidence values ​​are lower than a preset threshold, mainly by redetermining the category of the boundary node based on the label category of the neighboring nodes of each boundary node; Boundary nodes are determined according to the following principles: For a non-cell node u, if there exists a first-degree topological neighbor v (the parent or child node of u) in the tree topology graph, and v has a different label than u and v is not a cell node, then the non-cell node u is determined to be a boundary node.

[0103] Reclassification targets: Reclassification is performed only on boundary nodes with a confidence value less than a preset threshold (e.g., 0.7).

[0104] Calculate the centroid coordinates c of the axonal and dendritic compartments respectively. axon c dend (The arithmetic mean of the three-dimensional coordinates of all nodes within the zone).

[0105] d axon d dend Let represent the Euclidean distances from the non-cell body node u to be reclassified to the centroid of the axonal compartment and the centroid of the dendritic compartment, respectively. The specific calculation formulas are as follows: d axon =‖p u -c axon ||; d dend =‖p u -c dend ||; Where, p u The coordinates of the non-cell node u are given.

[0106] For a non-cell node u, count the labels of all its neighboring nodes.

[0107] Let N axon N is the number of neighboring nodes of a non-cell body node u whose labels are axon labels. dend The number of neighboring nodes of non-cell body node u, whose labels are dendritic labels.

[0108] Then the axonal score Scroe axon The calculation formula is as follows: Scroe axon =0.4×[1 / (1+d axon )]+0.6×[N axon / max(N axon +N dend ,1)]; Dendritic score Scroe dend The calculation formula is as follows: Scroe dend =0.4×[1 / (1+d dend )]+0.6×[N dend / max(N axon +N dend ,1)]; If Scroe axon >Scroe dend Then, update the label of node u to the axon label, and update the confidence value C of node u according to the following formula. u : C u =Scroe axon / (Scroe axon +Scroe dend ); If Scroe axon ≤Scroe dend Then, update the label of node u to the dendrite label, and update the confidence value C of node u according to the following formula. u : C u =Scroe dend / (Scroe axon +Scroe dend ).

[0109] S45, iterate through S43-S44 until the confidence values ​​of all nodes downstream of the main branch point converge, that is, the change in the confidence value of a node between adjacent iterations is less than the convergence threshold (which can be customized as needed, and is not limited here), and obtain the final classification result.

[0110] In step S5, the original SWC file is updated based on the final classification result of S45, and the updated SWC file is output. Label 2 is used to identify axonal nodes, label 3 to identify dendritic nodes, and label 1 remains unchanged for cell body nodes. The output file maintains the same coordinate system and topological connectivity as the original SWC file and can be directly used for subsequent morphological analysis and visualization.

[0111] The S5 process also outputs a quality assessment report. This report includes key metrics such as Precision and F1 score, providing users with quantitative indicators of the reliability of the classification results and facilitating the identification of low-confidence value regions that may require manual review. The specific methods for determining metrics such as Precision and F1 score are known techniques and not innovative aspects of this invention; therefore, they will not be elaborated upon here.

[0112] This invention has many other embodiments, and all technical solutions formed by equivalent transformation or equivalent transformation fall within the protection scope of this invention.

Claims

1. A method for classifying neuronal axons and dendrites, characterized in that, Includes the following steps: S1, Obtain the original SWC file, and generate a tree topology diagram of neurons based on the original SWC file; S2, determine a main branch point from the tree topology of the neuron; S3, determine the categories of the two main subtrees of the main branch point to obtain preliminary classification results; S4, refine the preliminary classification results based on the confidence value propagation algorithm to obtain the final classification results; S5 updates the original SWC file based on the final classification results and outputs the updated SWC file.

2. The neuron axon and dendrite classification method according to claim 1, characterized in that: S2 includes: S21, determine candidate branch points based on the tree topology diagram; S22, determine the subtree asymmetry score, spatial separation score, topological distance score and morphological difference score for each candidate branch point; S23. Calculate the comprehensive score of each candidate branch point based on the subtree asymmetry score, spatial separation score, topological distance score and morphological difference score of each candidate branch point. S24, the candidate branch with the highest comprehensive score is taken as the main branch.

3. The neuron axon and dendrite classification method according to claim 2, characterized in that: In S23, the comprehensive score of each candidate branch point is obtained by weighted summation of its subtree asymmetry score, spatial separation score, topological distance score, and morphological difference score.

4. The neuron axon and dendrite classification method according to claim 3, characterized in that: When determining the weights of subtree asymmetry score, spatial separation score, topological distance score, and morphological difference score for each candidate branch point, the global morphological features of the tree topology are first determined. Then, it is determined whether the weights need to be adjusted based on the global morphological features. If so, the weights are adjusted according to preset rules; otherwise, the default weights are used.

5. The neuron axon and dendrite classification method according to claim 4, characterized in that: The default weights for the subtree asymmetry score, spatial separation score, topological distance score, and morphological difference score are 0.4, 0.3, 0.2, and 0.1, respectively.

6. The neuron axon and dendrite classification method according to claim 3, characterized in that: The global morphological features of the tree topology include global branch ratio, global spatial extensibility, and global maximum topological depth.

7. The neuronal axon and dendrite classification method according to claim 1, characterized in that: S3 includes: S31, divide all direct subtrees of the main branch point into two mutually exclusive main subtrees; S32, determine the distal scale feature value, density feature value and orientation feature value of the two main subtrees respectively; S33, determine whether the absolute value of the difference between the distal scale feature values ​​of the two main subtrees exceeds a first set value; if yes, determine the main subtree with the larger distal scale feature value as an axon and the main subtree with the smaller distal scale feature value as a dendrite; if no, execute S34. S34, determine whether the absolute value of the difference between the density feature values ​​of the two main subtrees exceeds the second set value. If yes, determine the main subtree with the larger density feature value as a dendrite and the main subtree with the smaller density feature value as an axon. If no, proceed to S35. S35, determine whether the absolute value of the difference between the directional feature values ​​of the two main subtrees exceeds the third preset value. If yes, determine the main subtree with the higher directional feature value as an axon and the main subtree with the lower directional feature value as a dendrite. If no, proceed to S36. S36, the main subtree with more nodes in the two main subtrees is identified as an axon, and the main subtree with fewer nodes is identified as a dendrite.

8. The neuronal axon and dendrite classification method according to any one of claims 1-7, characterized in that: S4 includes: S41, perform anomaly detection on the preliminary classification results, and correct anomalies when they are detected; S42, assign initial confidence values ​​to the main branch point and all its downstream nodes; S43, Update the confidence values ​​of the main branch point and all its downstream nodes; S44, perform reclassification and update the confidence value for boundary nodes located on the axon and dendrite boundaries and whose confidence values ​​are lower than the preset threshold; S45, iterate through S43-S44 until the confidence values ​​of all nodes downstream of the main branch point converge, and obtain the final classification result.

9. The neuron axon and dendrite classification method according to claim 8, characterized in that: In step S43, the confidence value of each node is updated according to the following formula: The updated node confidence value = 0.7 × the current node confidence value + 0.3 × the support ratio of neighbors with the same label.

10. The neuron axon and dendrite classification method according to claim 1, characterized in that: The S5 also outputs a quality assessment report.